FFmpeg coverage


Directory: ../../../ffmpeg/
File: src/libavcodec/speexdec.c
Date: 2025-06-01 09:29:47
Exec Total Coverage
Lines: 99 776 12.8%
Functions: 6 31 19.4%
Branches: 32 456 7.0%

Line Branch Exec Source
1 /*
2 * Copyright 2002-2008 Xiph.org Foundation
3 * Copyright 2002-2008 Jean-Marc Valin
4 * Copyright 2005-2007 Analog Devices Inc.
5 * Copyright 2005-2008 Commonwealth Scientific and Industrial Research Organisation (CSIRO)
6 * Copyright 1993, 2002, 2006 David Rowe
7 * Copyright 2003 EpicGames
8 * Copyright 1992-1994 Jutta Degener, Carsten Bormann
9
10 * Redistribution and use in source and binary forms, with or without
11 * modification, are permitted provided that the following conditions
12 * are met:
13
14 * - Redistributions of source code must retain the above copyright
15 * notice, this list of conditions and the following disclaimer.
16
17 * - Redistributions in binary form must reproduce the above copyright
18 * notice, this list of conditions and the following disclaimer in the
19 * documentation and/or other materials provided with the distribution.
20
21 * - Neither the name of the Xiph.org Foundation nor the names of its
22 * contributors may be used to endorse or promote products derived from
23 * this software without specific prior written permission.
24
25 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
26 * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
27 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
28 * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE FOUNDATION OR
29 * CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
30 * EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
31 * PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
32 * PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF
33 * LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING
34 * NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
35 * SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
36 *
37 * This file is part of FFmpeg.
38 *
39 * FFmpeg is free software; you can redistribute it and/or
40 * modify it under the terms of the GNU Lesser General Public
41 * License as published by the Free Software Foundation; either
42 * version 2.1 of the License, or (at your option) any later version.
43 *
44 * FFmpeg is distributed in the hope that it will be useful,
45 * but WITHOUT ANY WARRANTY; without even the implied warranty of
46 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
47 * Lesser General Public License for more details.
48 *
49 * You should have received a copy of the GNU Lesser General Public
50 * License along with FFmpeg; if not, write to the Free Software
51 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
52 */
53
54 #include "libavutil/avassert.h"
55 #include "libavutil/avstring.h"
56 #include "libavutil/float_dsp.h"
57 #include "libavutil/intfloat.h"
58 #include "libavutil/mem.h"
59 #include "avcodec.h"
60 #include "bytestream.h"
61 #include "codec_internal.h"
62 #include "decode.h"
63 #include "get_bits.h"
64 #include "speexdata.h"
65
66 #define SPEEX_NB_MODES 3
67 #define SPEEX_INBAND_STEREO 9
68
69 #define QMF_ORDER 64
70 #define NB_ORDER 10
71 #define NB_FRAME_SIZE 160
72 #define NB_SUBMODES 9
73 #define NB_SUBMODE_BITS 4
74 #define SB_SUBMODE_BITS 3
75
76 #define NB_SUBFRAME_SIZE 40
77 #define NB_NB_SUBFRAMES 4
78 #define NB_PITCH_START 17
79 #define NB_PITCH_END 144
80
81 #define NB_DEC_BUFFER (NB_FRAME_SIZE + 2 * NB_PITCH_END + NB_SUBFRAME_SIZE + 12)
82
83 #define SPEEX_MEMSET(dst, c, n) (memset((dst), (c), (n) * sizeof(*(dst))))
84 #define SPEEX_COPY(dst, src, n) (memcpy((dst), (src), (n) * sizeof(*(dst))))
85
86 #define LSP_LINEAR(i) (.25f * (i) + .25f)
87 #define LSP_LINEAR_HIGH(i) (.3125f * (i) + .75f)
88 #define LSP_DIV_256(x) (0.00390625f * (x))
89 #define LSP_DIV_512(x) (0.001953125f * (x))
90 #define LSP_DIV_1024(x) (0.0009765625f * (x))
91
92 typedef struct LtpParams {
93 const int8_t *gain_cdbk;
94 int gain_bits;
95 int pitch_bits;
96 } LtpParam;
97
98 static const LtpParam ltp_params_vlbr = { gain_cdbk_lbr, 5, 0 };
99 static const LtpParam ltp_params_lbr = { gain_cdbk_lbr, 5, 7 };
100 static const LtpParam ltp_params_med = { gain_cdbk_lbr, 5, 7 };
101 static const LtpParam ltp_params_nb = { gain_cdbk_nb, 7, 7 };
102
103 typedef struct SplitCodebookParams {
104 int subvect_size;
105 int nb_subvect;
106 const signed char *shape_cb;
107 int shape_bits;
108 int have_sign;
109 } SplitCodebookParams;
110
111 static const SplitCodebookParams split_cb_nb_ulbr = { 20, 2, exc_20_32_table, 5, 0 };
112 static const SplitCodebookParams split_cb_nb_vlbr = { 10, 4, exc_10_16_table, 4, 0 };
113 static const SplitCodebookParams split_cb_nb_lbr = { 10, 4, exc_10_32_table, 5, 0 };
114 static const SplitCodebookParams split_cb_nb_med = { 8, 5, exc_8_128_table, 7, 0 };
115 static const SplitCodebookParams split_cb_nb = { 5, 8, exc_5_64_table, 6, 0 };
116 static const SplitCodebookParams split_cb_sb = { 5, 8, exc_5_256_table, 8, 0 };
117 static const SplitCodebookParams split_cb_high = { 8, 5, hexc_table, 7, 1 };
118 static const SplitCodebookParams split_cb_high_lbr= { 10, 4, hexc_10_32_table,5, 0 };
119
120 /** Quantizes LSPs */
121 typedef void (*lsp_quant_func)(float *, float *, int, GetBitContext *);
122
123 /** Decodes quantized LSPs */
124 typedef void (*lsp_unquant_func)(float *, int, GetBitContext *);
125
126 /** Long-term predictor quantization */
127 typedef int (*ltp_quant_func)(float *, float *, float *,
128 float *, float *, float *,
129 const void *, int, int, float, int, int,
130 GetBitContext *, char *, float *,
131 float *, int, int, int, float *);
132
133 /** Long-term un-quantize */
134 typedef void (*ltp_unquant_func)(float *, float *, int, int,
135 float, const void *, int, int *,
136 float *, GetBitContext *, int, int,
137 float, int);
138
139 /** Innovation quantization function */
140 typedef void (*innovation_quant_func)(float *, float *,
141 float *, float *, const void *,
142 int, int, float *, float *,
143 GetBitContext *, char *, int, int);
144
145 /** Innovation unquantization function */
146 typedef void (*innovation_unquant_func)(float *, const void *, int,
147 GetBitContext *, uint32_t *);
148
149 typedef struct SpeexSubmode {
150 int lbr_pitch; /**< Set to -1 for "normal" modes, otherwise encode pitch using
151 a global pitch and allowing a +- lbr_pitch variation (for
152 low not-rates)*/
153 int forced_pitch_gain; /**< Use the same (forced) pitch gain for all
154 sub-frames */
155 int have_subframe_gain; /**< Number of bits to use as sub-frame innovation
156 gain */
157 int double_codebook; /**< Apply innovation quantization twice for higher
158 quality (and higher bit-rate)*/
159 lsp_unquant_func lsp_unquant; /**< LSP unquantization function */
160
161 ltp_unquant_func ltp_unquant; /**< Long-term predictor (pitch) un-quantizer */
162 const void *LtpParam; /**< Pitch parameters (options) */
163
164 innovation_unquant_func innovation_unquant; /**< Innovation un-quantization */
165 const void *innovation_params; /**< Innovation quantization parameters*/
166
167 float comb_gain; /**< Gain of enhancer comb filter */
168 } SpeexSubmode;
169
170 typedef struct SpeexMode {
171 int modeID; /**< ID of the mode */
172 int (*decode)(AVCodecContext *avctx, void *dec, GetBitContext *gb, float *out, int packets_left);
173 int frame_size; /**< Size of frames used for decoding */
174 int subframe_size; /**< Size of sub-frames used for decoding */
175 int lpc_size; /**< Order of LPC filter */
176 float folding_gain; /**< Folding gain */
177 const SpeexSubmode *submodes[NB_SUBMODES]; /**< Sub-mode data for the mode */
178 int default_submode; /**< Default sub-mode to use when decoding */
179 } SpeexMode;
180
181 typedef struct DecoderState {
182 const SpeexMode *mode;
183 int modeID; /**< ID of the decoder mode */
184 int first; /**< Is first frame */
185 int full_frame_size; /**< Length of full-band frames */
186 int is_wideband; /**< If wideband is present */
187 int count_lost; /**< Was the last frame lost? */
188 int frame_size; /**< Length of high-band frames */
189 int subframe_size; /**< Length of high-band sub-frames */
190 int nb_subframes; /**< Number of high-band sub-frames */
191 int lpc_size; /**< Order of high-band LPC analysis */
192 float last_ol_gain; /**< Open-loop gain for previous frame */
193 float *innov_save; /**< If non-NULL, innovation is copied here */
194
195 /* This is used in packet loss concealment */
196 int last_pitch; /**< Pitch of last correctly decoded frame */
197 float last_pitch_gain; /**< Pitch gain of last correctly decoded frame */
198 uint32_t seed; /**< Seed used for random number generation */
199
200 int encode_submode;
201 const SpeexSubmode *const *submodes; /**< Sub-mode data */
202 int submodeID; /**< Activated sub-mode */
203 int lpc_enh_enabled; /**< 1 when LPC enhancer is on, 0 otherwise */
204
205 /* Vocoder data */
206 float voc_m1;
207 float voc_m2;
208 float voc_mean;
209 int voc_offset;
210
211 int dtx_enabled;
212 int highpass_enabled; /**< Is the input filter enabled */
213
214 float *exc; /**< Start of excitation frame */
215 float mem_hp[2]; /**< High-pass filter memory */
216 float exc_buf[NB_DEC_BUFFER]; /**< Excitation buffer */
217 float old_qlsp[NB_ORDER]; /**< Quantized LSPs for previous frame */
218 float interp_qlpc[NB_ORDER]; /**< Interpolated quantized LPCs */
219 float mem_sp[NB_ORDER]; /**< Filter memory for synthesis signal */
220 float g0_mem[QMF_ORDER];
221 float g1_mem[QMF_ORDER];
222 float pi_gain[NB_NB_SUBFRAMES]; /**< Gain of LPC filter at theta=pi (fe/2) */
223 float exc_rms[NB_NB_SUBFRAMES]; /**< RMS of excitation per subframe */
224 } DecoderState;
225
226 /* Default handler for user callbacks: skip it */
227 static int speex_default_user_handler(GetBitContext *gb, void *state, void *data)
228 {
229 const int req_size = get_bits(gb, 4);
230 skip_bits_long(gb, 5 + 8 * req_size);
231 return 0;
232 }
233
234 typedef struct StereoState {
235 float balance; /**< Left/right balance info */
236 float e_ratio; /**< Ratio of energies: E(left+right)/[E(left)+E(right)] */
237 float smooth_left; /**< Smoothed left channel gain */
238 float smooth_right; /**< Smoothed right channel gain */
239 } StereoState;
240
241 typedef struct SpeexContext {
242 AVClass *class;
243 GetBitContext gb;
244
245 int32_t version_id; /**< Version for Speex (for checking compatibility) */
246 int32_t rate; /**< Sampling rate used */
247 int32_t mode; /**< Mode used (0 for narrowband, 1 for wideband) */
248 int32_t bitstream_version; /**< Version ID of the bit-stream */
249 int32_t nb_channels; /**< Number of channels decoded */
250 int32_t bitrate; /**< Bit-rate used */
251 int32_t frame_size; /**< Size of frames */
252 int32_t vbr; /**< 1 for a VBR decoding, 0 otherwise */
253 int32_t frames_per_packet; /**< Number of frames stored per Ogg packet */
254 int32_t extra_headers; /**< Number of additional headers after the comments */
255
256 int pkt_size;
257
258 StereoState stereo;
259 DecoderState st[SPEEX_NB_MODES];
260
261 AVFloatDSPContext *fdsp;
262 } SpeexContext;
263
264 static void lsp_unquant_lbr(float *lsp, int order, GetBitContext *gb)
265 {
266 int id;
267
268 for (int i = 0; i < order; i++)
269 lsp[i] = LSP_LINEAR(i);
270
271 id = get_bits(gb, 6);
272 for (int i = 0; i < 10; i++)
273 lsp[i] += LSP_DIV_256(cdbk_nb[id * 10 + i]);
274
275 id = get_bits(gb, 6);
276 for (int i = 0; i < 5; i++)
277 lsp[i] += LSP_DIV_512(cdbk_nb_low1[id * 5 + i]);
278
279 id = get_bits(gb, 6);
280 for (int i = 0; i < 5; i++)
281 lsp[i + 5] += LSP_DIV_512(cdbk_nb_high1[id * 5 + i]);
282 }
283
284 static void forced_pitch_unquant(float *exc, float *exc_out, int start, int end,
285 float pitch_coef, const void *par, int nsf,
286 int *pitch_val, float *gain_val, GetBitContext *gb, int count_lost,
287 int subframe_offset, float last_pitch_gain, int cdbk_offset)
288 {
289 av_assert0(!isnan(pitch_coef));
290 pitch_coef = fminf(pitch_coef, .99f);
291 for (int i = 0; i < nsf; i++) {
292 exc_out[i] = exc[i - start] * pitch_coef;
293 exc[i] = exc_out[i];
294 }
295 pitch_val[0] = start;
296 gain_val[0] = gain_val[2] = 0.f;
297 gain_val[1] = pitch_coef;
298 }
299
300 static inline float speex_rand(float std, uint32_t *seed)
301 {
302 const uint32_t jflone = 0x3f800000;
303 const uint32_t jflmsk = 0x007fffff;
304 float fran;
305 uint32_t ran;
306 seed[0] = 1664525 * seed[0] + 1013904223;
307 ran = jflone | (jflmsk & seed[0]);
308 fran = av_int2float(ran);
309 fran -= 1.5f;
310 fran *= std;
311 return fran;
312 }
313
314 static void noise_codebook_unquant(float *exc, const void *par, int nsf,
315 GetBitContext *gb, uint32_t *seed)
316 {
317 for (int i = 0; i < nsf; i++)
318 exc[i] = speex_rand(1.f, seed);
319 }
320
321 static void split_cb_shape_sign_unquant(float *exc, const void *par, int nsf,
322 GetBitContext *gb, uint32_t *seed)
323 {
324 int subvect_size, nb_subvect, have_sign, shape_bits;
325 const SplitCodebookParams *params;
326 const signed char *shape_cb;
327 int signs[10], ind[10];
328
329 params = par;
330 subvect_size = params->subvect_size;
331 nb_subvect = params->nb_subvect;
332
333 shape_cb = params->shape_cb;
334 have_sign = params->have_sign;
335 shape_bits = params->shape_bits;
336
337 /* Decode codewords and gains */
338 for (int i = 0; i < nb_subvect; i++) {
339 signs[i] = have_sign ? get_bits1(gb) : 0;
340 ind[i] = get_bitsz(gb, shape_bits);
341 }
342 /* Compute decoded excitation */
343 for (int i = 0; i < nb_subvect; i++) {
344 const float s = signs[i] ? -1.f : 1.f;
345
346 for (int j = 0; j < subvect_size; j++)
347 exc[subvect_size * i + j] += s * 0.03125f * shape_cb[ind[i] * subvect_size + j];
348 }
349 }
350
351 #define SUBMODE(x) st->submodes[st->submodeID]->x
352
353 #define gain_3tap_to_1tap(g) (FFABS(g[1]) + (g[0] > 0.f ? g[0] : -.5f * g[0]) + (g[2] > 0.f ? g[2] : -.5f * g[2]))
354
355 static void
356 pitch_unquant_3tap(float *exc, float *exc_out, int start, int end, float pitch_coef,
357 const void *par, int nsf, int *pitch_val, float *gain_val, GetBitContext *gb,
358 int count_lost, int subframe_offset, float last_pitch_gain, int cdbk_offset)
359 {
360 int pitch, gain_index, gain_cdbk_size;
361 const int8_t *gain_cdbk;
362 const LtpParam *params;
363 float gain[3];
364
365 params = (const LtpParam *)par;
366 gain_cdbk_size = 1 << params->gain_bits;
367 gain_cdbk = params->gain_cdbk + 4 * gain_cdbk_size * cdbk_offset;
368
369 pitch = get_bitsz(gb, params->pitch_bits);
370 pitch += start;
371 gain_index = get_bitsz(gb, params->gain_bits);
372 gain[0] = 0.015625f * gain_cdbk[gain_index * 4] + .5f;
373 gain[1] = 0.015625f * gain_cdbk[gain_index * 4 + 1] + .5f;
374 gain[2] = 0.015625f * gain_cdbk[gain_index * 4 + 2] + .5f;
375
376 if (count_lost && pitch > subframe_offset) {
377 float tmp = count_lost < 4 ? last_pitch_gain : 0.5f * last_pitch_gain;
378 float gain_sum;
379
380 tmp = fminf(tmp, .95f);
381 gain_sum = gain_3tap_to_1tap(gain);
382
383 if (gain_sum > tmp && gain_sum > 0.f) {
384 float fact = tmp / gain_sum;
385 for (int i = 0; i < 3; i++)
386 gain[i] *= fact;
387 }
388 }
389
390 pitch_val[0] = pitch;
391 gain_val[0] = gain[0];
392 gain_val[1] = gain[1];
393 gain_val[2] = gain[2];
394 SPEEX_MEMSET(exc_out, 0, nsf);
395
396 for (int i = 0; i < 3; i++) {
397 int tmp1, tmp3;
398 int pp = pitch + 1 - i;
399 tmp1 = nsf;
400 if (tmp1 > pp)
401 tmp1 = pp;
402 for (int j = 0; j < tmp1; j++)
403 exc_out[j] += gain[2 - i] * exc[j - pp];
404 tmp3 = nsf;
405 if (tmp3 > pp + pitch)
406 tmp3 = pp + pitch;
407 for (int j = tmp1; j < tmp3; j++)
408 exc_out[j] += gain[2 - i] * exc[j - pp - pitch];
409 }
410 }
411
412 static void lsp_unquant_nb(float *lsp, int order, GetBitContext *gb)
413 {
414 int id;
415
416 for (int i = 0; i < order; i++)
417 lsp[i] = LSP_LINEAR(i);
418
419 id = get_bits(gb, 6);
420 for (int i = 0; i < 10; i++)
421 lsp[i] += LSP_DIV_256(cdbk_nb[id * 10 + i]);
422
423 id = get_bits(gb, 6);
424 for (int i = 0; i < 5; i++)
425 lsp[i] += LSP_DIV_512(cdbk_nb_low1[id * 5 + i]);
426
427 id = get_bits(gb, 6);
428 for (int i = 0; i < 5; i++)
429 lsp[i] += LSP_DIV_1024(cdbk_nb_low2[id * 5 + i]);
430
431 id = get_bits(gb, 6);
432 for (int i = 0; i < 5; i++)
433 lsp[i + 5] += LSP_DIV_512(cdbk_nb_high1[id * 5 + i]);
434
435 id = get_bits(gb, 6);
436 for (int i = 0; i < 5; i++)
437 lsp[i + 5] += LSP_DIV_1024(cdbk_nb_high2[id * 5 + i]);
438 }
439
440 static void lsp_unquant_high(float *lsp, int order, GetBitContext *gb)
441 {
442 int id;
443
444 for (int i = 0; i < order; i++)
445 lsp[i] = LSP_LINEAR_HIGH(i);
446
447 id = get_bits(gb, 6);
448 for (int i = 0; i < order; i++)
449 lsp[i] += LSP_DIV_256(high_lsp_cdbk[id * order + i]);
450
451 id = get_bits(gb, 6);
452 for (int i = 0; i < order; i++)
453 lsp[i] += LSP_DIV_512(high_lsp_cdbk2[id * order + i]);
454 }
455
456 /* 2150 bps "vocoder-like" mode for comfort noise */
457 static const SpeexSubmode nb_submode1 = {
458 0, 1, 0, 0, lsp_unquant_lbr, forced_pitch_unquant, NULL,
459 noise_codebook_unquant, NULL, -1.f
460 };
461
462 /* 5.95 kbps very low bit-rate mode */
463 static const SpeexSubmode nb_submode2 = {
464 0, 0, 0, 0, lsp_unquant_lbr, pitch_unquant_3tap, &ltp_params_vlbr,
465 split_cb_shape_sign_unquant, &split_cb_nb_vlbr, .6f
466 };
467
468 /* 8 kbps low bit-rate mode */
469 static const SpeexSubmode nb_submode3 = {
470 -1, 0, 1, 0, lsp_unquant_lbr, pitch_unquant_3tap, &ltp_params_lbr,
471 split_cb_shape_sign_unquant, &split_cb_nb_lbr, .55f
472 };
473
474 /* 11 kbps medium bit-rate mode */
475 static const SpeexSubmode nb_submode4 = {
476 -1, 0, 1, 0, lsp_unquant_lbr, pitch_unquant_3tap, &ltp_params_med,
477 split_cb_shape_sign_unquant, &split_cb_nb_med, .45f
478 };
479
480 /* 15 kbps high bit-rate mode */
481 static const SpeexSubmode nb_submode5 = {
482 -1, 0, 3, 0, lsp_unquant_nb, pitch_unquant_3tap, &ltp_params_nb,
483 split_cb_shape_sign_unquant, &split_cb_nb, .25f
484 };
485
486 /* 18.2 high bit-rate mode */
487 static const SpeexSubmode nb_submode6 = {
488 -1, 0, 3, 0, lsp_unquant_nb, pitch_unquant_3tap, &ltp_params_nb,
489 split_cb_shape_sign_unquant, &split_cb_sb, .15f
490 };
491
492 /* 24.6 kbps high bit-rate mode */
493 static const SpeexSubmode nb_submode7 = {
494 -1, 0, 3, 1, lsp_unquant_nb, pitch_unquant_3tap, &ltp_params_nb,
495 split_cb_shape_sign_unquant, &split_cb_nb, 0.05f
496 };
497
498 /* 3.95 kbps very low bit-rate mode */
499 static const SpeexSubmode nb_submode8 = {
500 0, 1, 0, 0, lsp_unquant_lbr, forced_pitch_unquant, NULL,
501 split_cb_shape_sign_unquant, &split_cb_nb_ulbr, .5f
502 };
503
504 static const SpeexSubmode wb_submode1 = {
505 0, 0, 1, 0, lsp_unquant_high, NULL, NULL,
506 NULL, NULL, -1.f
507 };
508
509 static const SpeexSubmode wb_submode2 = {
510 0, 0, 1, 0, lsp_unquant_high, NULL, NULL,
511 split_cb_shape_sign_unquant, &split_cb_high_lbr, -1.f
512 };
513
514 static const SpeexSubmode wb_submode3 = {
515 0, 0, 1, 0, lsp_unquant_high, NULL, NULL,
516 split_cb_shape_sign_unquant, &split_cb_high, -1.f
517 };
518
519 static const SpeexSubmode wb_submode4 = {
520 0, 0, 1, 1, lsp_unquant_high, NULL, NULL,
521 split_cb_shape_sign_unquant, &split_cb_high, -1.f
522 };
523
524 static int nb_decode(AVCodecContext *, void *, GetBitContext *, float *, int packets_left);
525 static int sb_decode(AVCodecContext *, void *, GetBitContext *, float *, int packets_left);
526
527 static const SpeexMode speex_modes[SPEEX_NB_MODES] = {
528 {
529 .modeID = 0,
530 .decode = nb_decode,
531 .frame_size = NB_FRAME_SIZE,
532 .subframe_size = NB_SUBFRAME_SIZE,
533 .lpc_size = NB_ORDER,
534 .submodes = {
535 NULL, &nb_submode1, &nb_submode2, &nb_submode3, &nb_submode4,
536 &nb_submode5, &nb_submode6, &nb_submode7, &nb_submode8
537 },
538 .default_submode = 5,
539 },
540 {
541 .modeID = 1,
542 .decode = sb_decode,
543 .frame_size = NB_FRAME_SIZE,
544 .subframe_size = NB_SUBFRAME_SIZE,
545 .lpc_size = 8,
546 .folding_gain = 0.9f,
547 .submodes = {
548 NULL, &wb_submode1, &wb_submode2, &wb_submode3, &wb_submode4
549 },
550 .default_submode = 3,
551 },
552 {
553 .modeID = 2,
554 .decode = sb_decode,
555 .frame_size = 320,
556 .subframe_size = 80,
557 .lpc_size = 8,
558 .folding_gain = 0.7f,
559 .submodes = {
560 NULL, &wb_submode1
561 },
562 .default_submode = 1,
563 },
564 };
565
566 static float compute_rms(const float *x, int len)
567 {
568 float sum = 0.f;
569
570 for (int i = 0; i < len; i++)
571 sum += x[i] * x[i];
572
573 av_assert0(len > 0);
574 return sqrtf(.1f + sum / len);
575 }
576
577 static void bw_lpc(float gamma, const float *lpc_in,
578 float *lpc_out, int order)
579 {
580 float tmp = gamma;
581
582 for (int i = 0; i < order; i++) {
583 lpc_out[i] = tmp * lpc_in[i];
584 tmp *= gamma;
585 }
586 }
587
588 static void iir_mem(const float *x, const float *den,
589 float *y, int N, int ord, float *mem)
590 {
591 for (int i = 0; i < N; i++) {
592 float yi = x[i] + mem[0];
593 float nyi = -yi;
594 for (int j = 0; j < ord - 1; j++)
595 mem[j] = mem[j + 1] + den[j] * nyi;
596 mem[ord - 1] = den[ord - 1] * nyi;
597 y[i] = yi;
598 }
599 }
600
601 static void highpass(const float *x, float *y, int len, float *mem, int wide)
602 {
603 static const float Pcoef[2][3] = {{ 1.00000f, -1.92683f, 0.93071f }, { 1.00000f, -1.97226f, 0.97332f } };
604 static const float Zcoef[2][3] = {{ 0.96446f, -1.92879f, 0.96446f }, { 0.98645f, -1.97277f, 0.98645f } };
605 const float *den, *num;
606
607 den = Pcoef[wide];
608 num = Zcoef[wide];
609 for (int i = 0; i < len; i++) {
610 float yi = num[0] * x[i] + mem[0];
611 mem[0] = mem[1] + num[1] * x[i] + -den[1] * yi;
612 mem[1] = num[2] * x[i] + -den[2] * yi;
613 y[i] = yi;
614 }
615 }
616
617 #define median3(a, b, c) \
618 ((a) < (b) ? ((b) < (c) ? (b) : ((a) < (c) ? (c) : (a))) \
619 : ((c) < (b) ? (b) : ((c) < (a) ? (c) : (a))))
620
621 static int speex_std_stereo(GetBitContext *gb, void *state, void *data)
622 {
623 StereoState *stereo = data;
624 float sign = get_bits1(gb) ? -1.f : 1.f;
625
626 stereo->balance = exp(sign * .25f * get_bits(gb, 5));
627 stereo->e_ratio = e_ratio_quant[get_bits(gb, 2)];
628
629 return 0;
630 }
631
632 static int speex_inband_handler(GetBitContext *gb, void *state, StereoState *stereo)
633 {
634 int id = get_bits(gb, 4);
635
636 if (id == SPEEX_INBAND_STEREO) {
637 return speex_std_stereo(gb, state, stereo);
638 } else {
639 int adv;
640
641 if (id < 2)
642 adv = 1;
643 else if (id < 8)
644 adv = 4;
645 else if (id < 10)
646 adv = 8;
647 else if (id < 12)
648 adv = 16;
649 else if (id < 14)
650 adv = 32;
651 else
652 adv = 64;
653 skip_bits_long(gb, adv);
654 }
655 return 0;
656 }
657
658 static void sanitize_values(float *vec, float min_val, float max_val, int len)
659 {
660 for (int i = 0; i < len; i++) {
661 if (!isnormal(vec[i]) || fabsf(vec[i]) < 1e-8f)
662 vec[i] = 0.f;
663 else
664 vec[i] = av_clipf(vec[i], min_val, max_val);
665 }
666 }
667
668 static void signal_mul(const float *x, float *y, float scale, int len)
669 {
670 for (int i = 0; i < len; i++)
671 y[i] = scale * x[i];
672 }
673
674 static float inner_prod(const float *x, const float *y, int len)
675 {
676 float sum = 0.f;
677
678 for (int i = 0; i < len; i += 8) {
679 float part = 0.f;
680 part += x[i + 0] * y[i + 0];
681 part += x[i + 1] * y[i + 1];
682 part += x[i + 2] * y[i + 2];
683 part += x[i + 3] * y[i + 3];
684 part += x[i + 4] * y[i + 4];
685 part += x[i + 5] * y[i + 5];
686 part += x[i + 6] * y[i + 6];
687 part += x[i + 7] * y[i + 7];
688 sum += part;
689 }
690
691 return sum;
692 }
693
694 static int interp_pitch(const float *exc, float *interp, int pitch, int len)
695 {
696 float corr[4][7], maxcorr;
697 int maxi, maxj;
698
699 for (int i = 0; i < 7; i++)
700 corr[0][i] = inner_prod(exc, exc - pitch - 3 + i, len);
701 for (int i = 0; i < 3; i++) {
702 for (int j = 0; j < 7; j++) {
703 int i1, i2;
704 float tmp = 0.f;
705
706 i1 = 3 - j;
707 if (i1 < 0)
708 i1 = 0;
709 i2 = 10 - j;
710 if (i2 > 7)
711 i2 = 7;
712 for (int k = i1; k < i2; k++)
713 tmp += shift_filt[i][k] * corr[0][j + k - 3];
714 corr[i + 1][j] = tmp;
715 }
716 }
717 maxi = maxj = 0;
718 maxcorr = corr[0][0];
719 for (int i = 0; i < 4; i++) {
720 for (int j = 0; j < 7; j++) {
721 if (corr[i][j] > maxcorr) {
722 maxcorr = corr[i][j];
723 maxi = i;
724 maxj = j;
725 }
726 }
727 }
728 for (int i = 0; i < len; i++) {
729 float tmp = 0.f;
730 if (maxi > 0.f) {
731 for (int k = 0; k < 7; k++)
732 tmp += exc[i - (pitch - maxj + 3) + k - 3] * shift_filt[maxi - 1][k];
733 } else {
734 tmp = exc[i - (pitch - maxj + 3)];
735 }
736 interp[i] = tmp;
737 }
738 return pitch - maxj + 3;
739 }
740
741 static void multicomb(const float *exc, float *new_exc, float *ak, int p, int nsf,
742 int pitch, int max_pitch, float comb_gain)
743 {
744 float old_ener, new_ener;
745 float iexc0_mag, iexc1_mag, exc_mag;
746 float iexc[4 * NB_SUBFRAME_SIZE];
747 float corr0, corr1, gain0, gain1;
748 float pgain1, pgain2;
749 float c1, c2, g1, g2;
750 float ngain, gg1, gg2;
751 int corr_pitch = pitch;
752
753 interp_pitch(exc, iexc, corr_pitch, 80);
754 if (corr_pitch > max_pitch)
755 interp_pitch(exc, iexc + nsf, 2 * corr_pitch, 80);
756 else
757 interp_pitch(exc, iexc + nsf, -corr_pitch, 80);
758
759 iexc0_mag = sqrtf(1000.f + inner_prod(iexc, iexc, nsf));
760 iexc1_mag = sqrtf(1000.f + inner_prod(iexc + nsf, iexc + nsf, nsf));
761 exc_mag = sqrtf(1.f + inner_prod(exc, exc, nsf));
762 corr0 = inner_prod(iexc, exc, nsf);
763 corr1 = inner_prod(iexc + nsf, exc, nsf);
764 if (corr0 > iexc0_mag * exc_mag)
765 pgain1 = 1.f;
766 else
767 pgain1 = (corr0 / exc_mag) / iexc0_mag;
768 if (corr1 > iexc1_mag * exc_mag)
769 pgain2 = 1.f;
770 else
771 pgain2 = (corr1 / exc_mag) / iexc1_mag;
772 gg1 = exc_mag / iexc0_mag;
773 gg2 = exc_mag / iexc1_mag;
774 if (comb_gain > 0.f) {
775 c1 = .4f * comb_gain + .07f;
776 c2 = .5f + 1.72f * (c1 - .07f);
777 } else {
778 c1 = c2 = 0.f;
779 }
780 g1 = 1.f - c2 * pgain1 * pgain1;
781 g2 = 1.f - c2 * pgain2 * pgain2;
782 g1 = fmaxf(g1, c1);
783 g2 = fmaxf(g2, c1);
784 g1 = c1 / g1;
785 g2 = c1 / g2;
786
787 if (corr_pitch > max_pitch) {
788 gain0 = .7f * g1 * gg1;
789 gain1 = .3f * g2 * gg2;
790 } else {
791 gain0 = .6f * g1 * gg1;
792 gain1 = .6f * g2 * gg2;
793 }
794 for (int i = 0; i < nsf; i++)
795 new_exc[i] = exc[i] + (gain0 * iexc[i]) + (gain1 * iexc[i + nsf]);
796 new_ener = compute_rms(new_exc, nsf);
797 old_ener = compute_rms(exc, nsf);
798
799 old_ener = fmaxf(old_ener, 1.f);
800 new_ener = fmaxf(new_ener, 1.f);
801 old_ener = fminf(old_ener, new_ener);
802 ngain = old_ener / new_ener;
803
804 for (int i = 0; i < nsf; i++)
805 new_exc[i] *= ngain;
806 }
807
808 static void lsp_interpolate(const float *old_lsp, const float *new_lsp,
809 float *lsp, int len, int subframe,
810 int nb_subframes, float margin)
811 {
812 const float tmp = (1.f + subframe) / nb_subframes;
813
814 for (int i = 0; i < len; i++) {
815 lsp[i] = (1.f - tmp) * old_lsp[i] + tmp * new_lsp[i];
816 lsp[i] = av_clipf(lsp[i], margin, M_PI - margin);
817 }
818 for (int i = 1; i < len - 1; i++) {
819 lsp[i] = fmaxf(lsp[i], lsp[i - 1] + margin);
820 if (lsp[i] > lsp[i + 1] - margin)
821 lsp[i] = .5f * (lsp[i] + lsp[i + 1] - margin);
822 }
823 }
824
825 static void lsp_to_lpc(const float *freq, float *ak, int lpcrdr)
826 {
827 float xout1, xout2, xin1, xin2;
828 float *pw, *n0;
829 float Wp[4 * NB_ORDER + 2] = { 0 };
830 float x_freq[NB_ORDER];
831 const int m = lpcrdr >> 1;
832
833 pw = Wp;
834
835 xin1 = xin2 = 1.f;
836
837 for (int i = 0; i < lpcrdr; i++)
838 x_freq[i] = -cosf(freq[i]);
839
840 /* reconstruct P(z) and Q(z) by cascading second order
841 * polynomials in form 1 - 2xz(-1) +z(-2), where x is the
842 * LSP coefficient
843 */
844 for (int j = 0; j <= lpcrdr; j++) {
845 int i2 = 0;
846 for (int i = 0; i < m; i++, i2 += 2) {
847 n0 = pw + (i * 4);
848 xout1 = xin1 + 2.f * x_freq[i2 ] * n0[0] + n0[1];
849 xout2 = xin2 + 2.f * x_freq[i2 + 1] * n0[2] + n0[3];
850 n0[1] = n0[0];
851 n0[3] = n0[2];
852 n0[0] = xin1;
853 n0[2] = xin2;
854 xin1 = xout1;
855 xin2 = xout2;
856 }
857 xout1 = xin1 + n0[4];
858 xout2 = xin2 - n0[5];
859 if (j > 0)
860 ak[j - 1] = (xout1 + xout2) * 0.5f;
861 n0[4] = xin1;
862 n0[5] = xin2;
863
864 xin1 = 0.f;
865 xin2 = 0.f;
866 }
867 }
868
869 static int nb_decode(AVCodecContext *avctx, void *ptr_st,
870 GetBitContext *gb, float *out, int packets_left)
871 {
872 DecoderState *st = ptr_st;
873 float ol_gain = 0, ol_pitch_coef = 0, best_pitch_gain = 0, pitch_average = 0;
874 int m, pitch, wideband, ol_pitch = 0, best_pitch = 40;
875 SpeexContext *s = avctx->priv_data;
876 float innov[NB_SUBFRAME_SIZE];
877 float exc32[NB_SUBFRAME_SIZE];
878 float interp_qlsp[NB_ORDER];
879 float qlsp[NB_ORDER];
880 float ak[NB_ORDER];
881 float pitch_gain[3] = { 0 };
882
883 st->exc = st->exc_buf + 2 * NB_PITCH_END + NB_SUBFRAME_SIZE + 6;
884
885 if (st->encode_submode) {
886 do { /* Search for next narrowband block (handle requests, skip wideband blocks) */
887 if (get_bits_left(gb) < 5)
888 return AVERROR_INVALIDDATA;
889 wideband = get_bits1(gb);
890 if (wideband) /* Skip wideband block (for compatibility) */ {
891 int submode, advance;
892
893 submode = get_bits(gb, SB_SUBMODE_BITS);
894 advance = wb_skip_table[submode];
895 advance -= SB_SUBMODE_BITS + 1;
896 if (advance < 0)
897 return AVERROR_INVALIDDATA;
898 skip_bits_long(gb, advance);
899
900 if (get_bits_left(gb) < 5)
901 return AVERROR_INVALIDDATA;
902 wideband = get_bits1(gb);
903 if (wideband) {
904 submode = get_bits(gb, SB_SUBMODE_BITS);
905 advance = wb_skip_table[submode];
906 advance -= SB_SUBMODE_BITS + 1;
907 if (advance < 0)
908 return AVERROR_INVALIDDATA;
909 skip_bits_long(gb, advance);
910 wideband = get_bits1(gb);
911 if (wideband) {
912 av_log(avctx, AV_LOG_ERROR, "more than two wideband layers found\n");
913 return AVERROR_INVALIDDATA;
914 }
915 }
916 }
917 if (get_bits_left(gb) < 4)
918 return AVERROR_INVALIDDATA;
919 m = get_bits(gb, 4);
920 if (m == 15) /* We found a terminator */ {
921 return AVERROR_INVALIDDATA;
922 } else if (m == 14) /* Speex in-band request */ {
923 int ret = speex_inband_handler(gb, st, &s->stereo);
924 if (ret)
925 return ret;
926 } else if (m == 13) /* User in-band request */ {
927 int ret = speex_default_user_handler(gb, st, NULL);
928 if (ret)
929 return ret;
930 } else if (m > 8) /* Invalid mode */ {
931 return AVERROR_INVALIDDATA;
932 }
933 } while (m > 8);
934
935 st->submodeID = m; /* Get the sub-mode that was used */
936 }
937
938 /* Shift all buffers by one frame */
939 memmove(st->exc_buf, st->exc_buf + NB_FRAME_SIZE, (2 * NB_PITCH_END + NB_SUBFRAME_SIZE + 12) * sizeof(float));
940
941 /* If null mode (no transmission), just set a couple things to zero */
942 if (st->submodes[st->submodeID] == NULL) {
943 float lpc[NB_ORDER];
944 float innov_gain = 0.f;
945
946 bw_lpc(0.93f, st->interp_qlpc, lpc, NB_ORDER);
947 innov_gain = compute_rms(st->exc, NB_FRAME_SIZE);
948 for (int i = 0; i < NB_FRAME_SIZE; i++)
949 st->exc[i] = speex_rand(innov_gain, &st->seed);
950
951 /* Final signal synthesis from excitation */
952 iir_mem(st->exc, lpc, out, NB_FRAME_SIZE, NB_ORDER, st->mem_sp);
953 st->count_lost = 0;
954
955 return 0;
956 }
957
958 /* Unquantize LSPs */
959 SUBMODE(lsp_unquant)(qlsp, NB_ORDER, gb);
960
961 /* Damp memory if a frame was lost and the LSP changed too much */
962 if (st->count_lost) {
963 float fact, lsp_dist = 0;
964
965 for (int i = 0; i < NB_ORDER; i++)
966 lsp_dist = lsp_dist + FFABS(st->old_qlsp[i] - qlsp[i]);
967 fact = .6f * exp(-.2f * lsp_dist);
968 for (int i = 0; i < NB_ORDER; i++)
969 st->mem_sp[i] = fact * st->mem_sp[i];
970 }
971
972 /* Handle first frame and lost-packet case */
973 if (st->first || st->count_lost)
974 memcpy(st->old_qlsp, qlsp, sizeof(st->old_qlsp));
975
976 /* Get open-loop pitch estimation for low bit-rate pitch coding */
977 if (SUBMODE(lbr_pitch) != -1)
978 ol_pitch = NB_PITCH_START + get_bits(gb, 7);
979
980 if (SUBMODE(forced_pitch_gain))
981 ol_pitch_coef = 0.066667f * get_bits(gb, 4);
982
983 /* Get global excitation gain */
984 ol_gain = expf(get_bits(gb, 5) / 3.5f);
985
986 if (st->submodeID == 1)
987 st->dtx_enabled = get_bits(gb, 4) == 15;
988
989 if (st->submodeID > 1)
990 st->dtx_enabled = 0;
991
992 for (int sub = 0; sub < NB_NB_SUBFRAMES; sub++) { /* Loop on subframes */
993 float *exc, *innov_save = NULL, tmp, ener;
994 int pit_min, pit_max, offset, q_energy;
995
996 offset = NB_SUBFRAME_SIZE * sub; /* Offset relative to start of frame */
997 exc = st->exc + offset; /* Excitation */
998 if (st->innov_save) /* Original signal */
999 innov_save = st->innov_save + offset;
1000
1001 SPEEX_MEMSET(exc, 0, NB_SUBFRAME_SIZE); /* Reset excitation */
1002
1003 /* Adaptive codebook contribution */
1004 av_assert0(SUBMODE(ltp_unquant));
1005 /* Handle pitch constraints if any */
1006 if (SUBMODE(lbr_pitch) != -1) {
1007 int margin = SUBMODE(lbr_pitch);
1008
1009 if (margin) {
1010 pit_min = ol_pitch - margin + 1;
1011 pit_min = FFMAX(pit_min, NB_PITCH_START);
1012 pit_max = ol_pitch + margin;
1013 pit_max = FFMIN(pit_max, NB_PITCH_START);
1014 } else {
1015 pit_min = pit_max = ol_pitch;
1016 }
1017 } else {
1018 pit_min = NB_PITCH_START;
1019 pit_max = NB_PITCH_END;
1020 }
1021
1022 SUBMODE(ltp_unquant)(exc, exc32, pit_min, pit_max, ol_pitch_coef, SUBMODE(LtpParam),
1023 NB_SUBFRAME_SIZE, &pitch, pitch_gain, gb, st->count_lost, offset,
1024 st->last_pitch_gain, 0);
1025
1026 sanitize_values(exc32, -32000, 32000, NB_SUBFRAME_SIZE);
1027
1028 tmp = gain_3tap_to_1tap(pitch_gain);
1029
1030 pitch_average += tmp;
1031 if ((tmp > best_pitch_gain &&
1032 FFABS(2 * best_pitch - pitch) >= 3 &&
1033 FFABS(3 * best_pitch - pitch) >= 4 &&
1034 FFABS(4 * best_pitch - pitch) >= 5) ||
1035 (tmp > .6f * best_pitch_gain &&
1036 (FFABS(best_pitch - 2 * pitch) < 3 ||
1037 FFABS(best_pitch - 3 * pitch) < 4 ||
1038 FFABS(best_pitch - 4 * pitch) < 5)) ||
1039 ((.67f * tmp) > best_pitch_gain &&
1040 (FFABS(2 * best_pitch - pitch) < 3 ||
1041 FFABS(3 * best_pitch - pitch) < 4 ||
1042 FFABS(4 * best_pitch - pitch) < 5))) {
1043 best_pitch = pitch;
1044 if (tmp > best_pitch_gain)
1045 best_pitch_gain = tmp;
1046 }
1047
1048 memset(innov, 0, sizeof(innov));
1049
1050 /* Decode sub-frame gain correction */
1051 if (SUBMODE(have_subframe_gain) == 3) {
1052 q_energy = get_bits(gb, 3);
1053 ener = exc_gain_quant_scal3[q_energy] * ol_gain;
1054 } else if (SUBMODE(have_subframe_gain) == 1) {
1055 q_energy = get_bits1(gb);
1056 ener = exc_gain_quant_scal1[q_energy] * ol_gain;
1057 } else {
1058 ener = ol_gain;
1059 }
1060
1061 av_assert0(SUBMODE(innovation_unquant));
1062 /* Fixed codebook contribution */
1063 SUBMODE(innovation_unquant)(innov, SUBMODE(innovation_params), NB_SUBFRAME_SIZE, gb, &st->seed);
1064 /* De-normalize innovation and update excitation */
1065
1066 signal_mul(innov, innov, ener, NB_SUBFRAME_SIZE);
1067
1068 /* Decode second codebook (only for some modes) */
1069 if (SUBMODE(double_codebook)) {
1070 float innov2[NB_SUBFRAME_SIZE] = { 0 };
1071
1072 SUBMODE(innovation_unquant)(innov2, SUBMODE(innovation_params), NB_SUBFRAME_SIZE, gb, &st->seed);
1073 signal_mul(innov2, innov2, 0.454545f * ener, NB_SUBFRAME_SIZE);
1074 for (int i = 0; i < NB_SUBFRAME_SIZE; i++)
1075 innov[i] += innov2[i];
1076 }
1077 for (int i = 0; i < NB_SUBFRAME_SIZE; i++)
1078 exc[i] = exc32[i] + innov[i];
1079 if (innov_save)
1080 memcpy(innov_save, innov, sizeof(innov));
1081
1082 /* Vocoder mode */
1083 if (st->submodeID == 1) {
1084 float g = ol_pitch_coef;
1085
1086 g = av_clipf(1.5f * (g - .2f), 0.f, 1.f);
1087
1088 SPEEX_MEMSET(exc, 0, NB_SUBFRAME_SIZE);
1089 while (st->voc_offset < NB_SUBFRAME_SIZE) {
1090 if (st->voc_offset >= 0)
1091 exc[st->voc_offset] = sqrtf(2.f * ol_pitch) * (g * ol_gain);
1092 st->voc_offset += ol_pitch;
1093 }
1094 st->voc_offset -= NB_SUBFRAME_SIZE;
1095
1096 for (int i = 0; i < NB_SUBFRAME_SIZE; i++) {
1097 float exci = exc[i];
1098 exc[i] = (.7f * exc[i] + .3f * st->voc_m1) + ((1.f - .85f * g) * innov[i]) - .15f * g * st->voc_m2;
1099 st->voc_m1 = exci;
1100 st->voc_m2 = innov[i];
1101 st->voc_mean = .8f * st->voc_mean + .2f * exc[i];
1102 exc[i] -= st->voc_mean;
1103 }
1104 }
1105 }
1106
1107 if (st->lpc_enh_enabled && SUBMODE(comb_gain) > 0 && !st->count_lost) {
1108 multicomb(st->exc - NB_SUBFRAME_SIZE, out, st->interp_qlpc, NB_ORDER,
1109 2 * NB_SUBFRAME_SIZE, best_pitch, 40, SUBMODE(comb_gain));
1110 multicomb(st->exc + NB_SUBFRAME_SIZE, out + 2 * NB_SUBFRAME_SIZE,
1111 st->interp_qlpc, NB_ORDER, 2 * NB_SUBFRAME_SIZE, best_pitch, 40,
1112 SUBMODE(comb_gain));
1113 } else {
1114 SPEEX_COPY(out, &st->exc[-NB_SUBFRAME_SIZE], NB_FRAME_SIZE);
1115 }
1116
1117 /* If the last packet was lost, re-scale the excitation to obtain the same
1118 * energy as encoded in ol_gain */
1119 if (st->count_lost) {
1120 float exc_ener, gain;
1121
1122 exc_ener = compute_rms(st->exc, NB_FRAME_SIZE);
1123 av_assert0(exc_ener + 1.f > 0.f);
1124 gain = fminf(ol_gain / (exc_ener + 1.f), 2.f);
1125 for (int i = 0; i < NB_FRAME_SIZE; i++) {
1126 st->exc[i] *= gain;
1127 out[i] = st->exc[i - NB_SUBFRAME_SIZE];
1128 }
1129 }
1130
1131 for (int sub = 0; sub < NB_NB_SUBFRAMES; sub++) { /* Loop on subframes */
1132 const int offset = NB_SUBFRAME_SIZE * sub; /* Offset relative to start of frame */
1133 float pi_g = 1.f, *sp = out + offset; /* Original signal */
1134
1135 lsp_interpolate(st->old_qlsp, qlsp, interp_qlsp, NB_ORDER, sub, NB_NB_SUBFRAMES, 0.002f);
1136 lsp_to_lpc(interp_qlsp, ak, NB_ORDER); /* Compute interpolated LPCs (unquantized) */
1137
1138 for (int i = 0; i < NB_ORDER; i += 2) /* Compute analysis filter at w=pi */
1139 pi_g += ak[i + 1] - ak[i];
1140 st->pi_gain[sub] = pi_g;
1141 st->exc_rms[sub] = compute_rms(st->exc + offset, NB_SUBFRAME_SIZE);
1142
1143 iir_mem(sp, st->interp_qlpc, sp, NB_SUBFRAME_SIZE, NB_ORDER, st->mem_sp);
1144
1145 memcpy(st->interp_qlpc, ak, sizeof(st->interp_qlpc));
1146 }
1147
1148 if (st->highpass_enabled)
1149 highpass(out, out, NB_FRAME_SIZE, st->mem_hp, st->is_wideband);
1150
1151 /* Store the LSPs for interpolation in the next frame */
1152 memcpy(st->old_qlsp, qlsp, sizeof(st->old_qlsp));
1153
1154 st->count_lost = 0;
1155 st->last_pitch = best_pitch;
1156 st->last_pitch_gain = .25f * pitch_average;
1157 st->last_ol_gain = ol_gain;
1158 st->first = 0;
1159
1160 return 0;
1161 }
1162
1163 static void qmf_synth(const float *x1, const float *x2, const float *a, float *y, int N, int M, float *mem1, float *mem2)
1164 {
1165 const int M2 = M >> 1, N2 = N >> 1;
1166 float xx1[352], xx2[352];
1167
1168 for (int i = 0; i < N2; i++)
1169 xx1[i] = x1[N2-1-i];
1170 for (int i = 0; i < M2; i++)
1171 xx1[N2+i] = mem1[2*i+1];
1172 for (int i = 0; i < N2; i++)
1173 xx2[i] = x2[N2-1-i];
1174 for (int i = 0; i < M2; i++)
1175 xx2[N2+i] = mem2[2*i+1];
1176
1177 for (int i = 0; i < N2; i += 2) {
1178 float y0, y1, y2, y3;
1179 float x10, x20;
1180
1181 y0 = y1 = y2 = y3 = 0.f;
1182 x10 = xx1[N2-2-i];
1183 x20 = xx2[N2-2-i];
1184
1185 for (int j = 0; j < M2; j += 2) {
1186 float x11, x21;
1187 float a0, a1;
1188
1189 a0 = a[2*j];
1190 a1 = a[2*j+1];
1191 x11 = xx1[N2-1+j-i];
1192 x21 = xx2[N2-1+j-i];
1193
1194 y0 += a0 * (x11-x21);
1195 y1 += a1 * (x11+x21);
1196 y2 += a0 * (x10-x20);
1197 y3 += a1 * (x10+x20);
1198 a0 = a[2*j+2];
1199 a1 = a[2*j+3];
1200 x10 = xx1[N2+j-i];
1201 x20 = xx2[N2+j-i];
1202
1203 y0 += a0 * (x10-x20);
1204 y1 += a1 * (x10+x20);
1205 y2 += a0 * (x11-x21);
1206 y3 += a1 * (x11+x21);
1207 }
1208 y[2 * i ] = 2.f * y0;
1209 y[2 * i+1] = 2.f * y1;
1210 y[2 * i+2] = 2.f * y2;
1211 y[2 * i+3] = 2.f * y3;
1212 }
1213
1214 for (int i = 0; i < M2; i++)
1215 mem1[2*i+1] = xx1[i];
1216 for (int i = 0; i < M2; i++)
1217 mem2[2*i+1] = xx2[i];
1218 }
1219
1220 3 static int sb_decode(AVCodecContext *avctx, void *ptr_st,
1221 GetBitContext *gb, float *out, int packets_left)
1222 {
1223 3 SpeexContext *s = avctx->priv_data;
1224 3 DecoderState *st = ptr_st;
1225 float low_pi_gain[NB_NB_SUBFRAMES];
1226 float low_exc_rms[NB_NB_SUBFRAMES];
1227 float interp_qlsp[NB_ORDER];
1228 int ret, wideband;
1229 float *low_innov_alias;
1230 float qlsp[NB_ORDER];
1231 float ak[NB_ORDER];
1232 const SpeexMode *mode;
1233
1234 3 mode = st->mode;
1235
1236
1/2
✓ Branch 0 taken 3 times.
✗ Branch 1 not taken.
3 if (st->modeID > 0) {
1237
1/2
✓ Branch 0 taken 3 times.
✗ Branch 1 not taken.
3 if (packets_left <= 1)
1238 3 return AVERROR_INVALIDDATA;
1239 low_innov_alias = out + st->frame_size;
1240 s->st[st->modeID - 1].innov_save = low_innov_alias;
1241 ret = speex_modes[st->modeID - 1].decode(avctx, &s->st[st->modeID - 1], gb, out, packets_left);
1242 if (ret < 0)
1243 return ret;
1244 }
1245
1246 if (st->encode_submode) { /* Check "wideband bit" */
1247 if (get_bits_left(gb) > 0)
1248 wideband = show_bits1(gb);
1249 else
1250 wideband = 0;
1251 if (wideband) { /* Regular wideband frame, read the submode */
1252 wideband = get_bits1(gb);
1253 st->submodeID = get_bits(gb, SB_SUBMODE_BITS);
1254 } else { /* Was a narrowband frame, set "null submode" */
1255 st->submodeID = 0;
1256 }
1257 if (st->submodeID != 0 && st->submodes[st->submodeID] == NULL)
1258 return AVERROR_INVALIDDATA;
1259 }
1260
1261 /* If null mode (no transmission), just set a couple things to zero */
1262 if (st->submodes[st->submodeID] == NULL) {
1263 for (int i = 0; i < st->frame_size; i++)
1264 out[st->frame_size + i] = 1e-15f;
1265
1266 st->first = 1;
1267
1268 /* Final signal synthesis from excitation */
1269 iir_mem(out + st->frame_size, st->interp_qlpc, out + st->frame_size, st->frame_size, st->lpc_size, st->mem_sp);
1270
1271 qmf_synth(out, out + st->frame_size, h0, out, st->full_frame_size, QMF_ORDER, st->g0_mem, st->g1_mem);
1272
1273 return 0;
1274 }
1275
1276 memcpy(low_pi_gain, s->st[st->modeID - 1].pi_gain, sizeof(low_pi_gain));
1277 memcpy(low_exc_rms, s->st[st->modeID - 1].exc_rms, sizeof(low_exc_rms));
1278
1279 SUBMODE(lsp_unquant)(qlsp, st->lpc_size, gb);
1280
1281 if (st->first)
1282 memcpy(st->old_qlsp, qlsp, sizeof(st->old_qlsp));
1283
1284 for (int sub = 0; sub < st->nb_subframes; sub++) {
1285 float filter_ratio, el, rl, rh;
1286 float *innov_save = NULL, *sp;
1287 float exc[80];
1288 int offset;
1289
1290 offset = st->subframe_size * sub;
1291 sp = out + st->frame_size + offset;
1292 /* Pointer for saving innovation */
1293 if (st->innov_save) {
1294 innov_save = st->innov_save + 2 * offset;
1295 SPEEX_MEMSET(innov_save, 0, 2 * st->subframe_size);
1296 }
1297
1298 av_assert0(st->nb_subframes > 0);
1299 lsp_interpolate(st->old_qlsp, qlsp, interp_qlsp, st->lpc_size, sub, st->nb_subframes, 0.05f);
1300 lsp_to_lpc(interp_qlsp, ak, st->lpc_size);
1301
1302 /* Calculate reponse ratio between the low and high filter in the middle
1303 of the band (4000 Hz) */
1304 st->pi_gain[sub] = 1.f;
1305 rh = 1.f;
1306 for (int i = 0; i < st->lpc_size; i += 2) {
1307 rh += ak[i + 1] - ak[i];
1308 st->pi_gain[sub] += ak[i] + ak[i + 1];
1309 }
1310
1311 rl = low_pi_gain[sub];
1312 filter_ratio = (rl + .01f) / (rh + .01f);
1313
1314 SPEEX_MEMSET(exc, 0, st->subframe_size);
1315 if (!SUBMODE(innovation_unquant)) {
1316 const int x = get_bits(gb, 5);
1317 const float g = expf(.125f * (x - 10)) / filter_ratio;
1318
1319 for (int i = 0; i < st->subframe_size; i += 2) {
1320 exc[i ] = mode->folding_gain * low_innov_alias[offset + i ] * g;
1321 exc[i + 1] = -mode->folding_gain * low_innov_alias[offset + i + 1] * g;
1322 }
1323 } else {
1324 float gc, scale;
1325
1326 el = low_exc_rms[sub];
1327 gc = 0.87360f * gc_quant_bound[get_bits(gb, 4)];
1328
1329 if (st->subframe_size == 80)
1330 gc *= M_SQRT2;
1331
1332 scale = (gc * el) / filter_ratio;
1333 SUBMODE(innovation_unquant)
1334 (exc, SUBMODE(innovation_params), st->subframe_size,
1335 gb, &st->seed);
1336
1337 signal_mul(exc, exc, scale, st->subframe_size);
1338 if (SUBMODE(double_codebook)) {
1339 float innov2[80];
1340
1341 SPEEX_MEMSET(innov2, 0, st->subframe_size);
1342 SUBMODE(innovation_unquant)(innov2, SUBMODE(innovation_params), st->subframe_size, gb, &st->seed);
1343 signal_mul(innov2, innov2, 0.4f * scale, st->subframe_size);
1344 for (int i = 0; i < st->subframe_size; i++)
1345 exc[i] += innov2[i];
1346 }
1347 }
1348
1349 if (st->innov_save) {
1350 for (int i = 0; i < st->subframe_size; i++)
1351 innov_save[2 * i] = exc[i];
1352 }
1353
1354 iir_mem(st->exc_buf, st->interp_qlpc, sp, st->subframe_size, st->lpc_size, st->mem_sp);
1355 memcpy(st->exc_buf, exc, sizeof(exc));
1356 memcpy(st->interp_qlpc, ak, sizeof(st->interp_qlpc));
1357 st->exc_rms[sub] = compute_rms(st->exc_buf, st->subframe_size);
1358 }
1359
1360 qmf_synth(out, out + st->frame_size, h0, out, st->full_frame_size, QMF_ORDER, st->g0_mem, st->g1_mem);
1361 memcpy(st->old_qlsp, qlsp, sizeof(st->old_qlsp));
1362
1363 st->first = 0;
1364
1365 return 0;
1366 }
1367
1368 9 static int decoder_init(SpeexContext *s, DecoderState *st, const SpeexMode *mode)
1369 {
1370 9 st->mode = mode;
1371 9 st->modeID = mode->modeID;
1372
1373 9 st->first = 1;
1374 9 st->encode_submode = 1;
1375 9 st->is_wideband = st->modeID > 0;
1376 9 st->innov_save = NULL;
1377
1378 9 st->submodes = mode->submodes;
1379 9 st->submodeID = mode->default_submode;
1380 9 st->subframe_size = mode->subframe_size;
1381 9 st->lpc_size = mode->lpc_size;
1382
2/2
✓ Branch 0 taken 6 times.
✓ Branch 1 taken 3 times.
9 st->full_frame_size = (1 + (st->modeID > 0)) * mode->frame_size;
1383 9 st->nb_subframes = mode->frame_size / mode->subframe_size;
1384 9 st->frame_size = mode->frame_size;
1385
1386 9 st->lpc_enh_enabled = 1;
1387
1388 9 st->last_pitch = 40;
1389 9 st->count_lost = 0;
1390 9 st->seed = 1000;
1391 9 st->last_ol_gain = 0;
1392
1393 9 st->voc_m1 = st->voc_m2 = st->voc_mean = 0;
1394 9 st->voc_offset = 0;
1395 9 st->dtx_enabled = 0;
1396 9 st->highpass_enabled = mode->modeID == 0;
1397
1398 9 return 0;
1399 }
1400
1401 3 static int parse_speex_extradata(AVCodecContext *avctx,
1402 const uint8_t *extradata, int extradata_size)
1403 {
1404 3 SpeexContext *s = avctx->priv_data;
1405 3 const uint8_t *buf = av_strnstr(extradata, "Speex ", extradata_size);
1406
1407
1/2
✗ Branch 0 not taken.
✓ Branch 1 taken 3 times.
3 if (!buf)
1408 return AVERROR_INVALIDDATA;
1409
1410 3 buf += 28;
1411
1412 3 s->version_id = bytestream_get_le32(&buf);
1413 3 buf += 4;
1414 3 s->rate = bytestream_get_le32(&buf);
1415
1/2
✗ Branch 0 not taken.
✓ Branch 1 taken 3 times.
3 if (s->rate <= 0)
1416 return AVERROR_INVALIDDATA;
1417 3 s->mode = bytestream_get_le32(&buf);
1418
2/4
✓ Branch 0 taken 3 times.
✗ Branch 1 not taken.
✗ Branch 2 not taken.
✓ Branch 3 taken 3 times.
3 if (s->mode < 0 || s->mode >= SPEEX_NB_MODES)
1419 return AVERROR_INVALIDDATA;
1420 3 s->bitstream_version = bytestream_get_le32(&buf);
1421
1/2
✗ Branch 0 not taken.
✓ Branch 1 taken 3 times.
3 if (s->bitstream_version != 4)
1422 return AVERROR_INVALIDDATA;
1423 3 s->nb_channels = bytestream_get_le32(&buf);
1424
2/4
✓ Branch 0 taken 3 times.
✗ Branch 1 not taken.
✗ Branch 2 not taken.
✓ Branch 3 taken 3 times.
3 if (s->nb_channels <= 0 || s->nb_channels > 2)
1425 return AVERROR_INVALIDDATA;
1426 3 s->bitrate = bytestream_get_le32(&buf);
1427 3 s->frame_size = bytestream_get_le32(&buf);
1428
3/6
✓ Branch 0 taken 3 times.
✗ Branch 1 not taken.
✓ Branch 2 taken 3 times.
✗ Branch 3 not taken.
✗ Branch 4 not taken.
✓ Branch 5 taken 3 times.
6 if (s->frame_size < NB_FRAME_SIZE << (s->mode > 1) ||
1429
1/2
✓ Branch 0 taken 3 times.
✗ Branch 1 not taken.
3 s->frame_size > INT32_MAX >> (s->mode > 1))
1430 return AVERROR_INVALIDDATA;
1431 3 s->frame_size = FFMIN(s->frame_size << (s->mode > 1), NB_FRAME_SIZE << s->mode);
1432 3 s->vbr = bytestream_get_le32(&buf);
1433 3 s->frames_per_packet = bytestream_get_le32(&buf);
1434
1/2
✓ Branch 0 taken 3 times.
✗ Branch 1 not taken.
3 if (s->frames_per_packet <= 0 ||
1435
1/2
✓ Branch 0 taken 3 times.
✗ Branch 1 not taken.
3 s->frames_per_packet > 64 ||
1436
1/2
✗ Branch 0 not taken.
✓ Branch 1 taken 3 times.
3 s->frames_per_packet >= INT32_MAX / s->nb_channels / s->frame_size)
1437 return AVERROR_INVALIDDATA;
1438 3 s->extra_headers = bytestream_get_le32(&buf);
1439
1440 3 return 0;
1441 }
1442
1443 3 static av_cold int speex_decode_init(AVCodecContext *avctx)
1444 {
1445 3 SpeexContext *s = avctx->priv_data;
1446 int ret;
1447
1448 3 s->fdsp = avpriv_float_dsp_alloc(0);
1449
1/2
✗ Branch 0 not taken.
✓ Branch 1 taken 3 times.
3 if (!s->fdsp)
1450 return AVERROR(ENOMEM);
1451
1452
2/4
✓ Branch 0 taken 3 times.
✗ Branch 1 not taken.
✓ Branch 2 taken 3 times.
✗ Branch 3 not taken.
3 if (avctx->extradata && avctx->extradata_size >= 80) {
1453 3 ret = parse_speex_extradata(avctx, avctx->extradata, avctx->extradata_size);
1454
1/2
✗ Branch 0 not taken.
✓ Branch 1 taken 3 times.
3 if (ret < 0)
1455 return ret;
1456 } else {
1457 s->rate = avctx->sample_rate;
1458 if (s->rate <= 0)
1459 return AVERROR_INVALIDDATA;
1460
1461 s->nb_channels = avctx->ch_layout.nb_channels;
1462 if (s->nb_channels <= 0 || s->nb_channels > 2)
1463 return AVERROR_INVALIDDATA;
1464
1465 switch (s->rate) {
1466 case 8000: s->mode = 0; break;
1467 case 16000: s->mode = 1; break;
1468 case 32000: s->mode = 2; break;
1469 default: s->mode = 2;
1470 }
1471
1472 s->frames_per_packet = 64;
1473 s->frame_size = NB_FRAME_SIZE << s->mode;
1474 }
1475
1476
1/2
✗ Branch 0 not taken.
✓ Branch 1 taken 3 times.
3 if (avctx->codec_tag == MKTAG('S', 'P', 'X', 'N')) {
1477 int quality;
1478
1479 if (!avctx->extradata || avctx->extradata && avctx->extradata_size < 47) {
1480 av_log(avctx, AV_LOG_ERROR, "Missing or invalid extradata.\n");
1481 return AVERROR_INVALIDDATA;
1482 }
1483
1484 quality = avctx->extradata[37];
1485 if (quality > 10) {
1486 av_log(avctx, AV_LOG_ERROR, "Unsupported quality mode %d.\n", quality);
1487 return AVERROR_PATCHWELCOME;
1488 }
1489
1490 s->pkt_size = ((const uint8_t[]){ 5, 10, 15, 20, 20, 28, 28, 38, 38, 46, 62 })[quality];
1491
1492 s->mode = 0;
1493 s->nb_channels = 1;
1494 s->rate = avctx->sample_rate;
1495 if (s->rate <= 0)
1496 return AVERROR_INVALIDDATA;
1497 s->frames_per_packet = 1;
1498 s->frame_size = NB_FRAME_SIZE;
1499 }
1500
1501
1/2
✓ Branch 0 taken 3 times.
✗ Branch 1 not taken.
3 if (s->bitrate > 0)
1502 3 avctx->bit_rate = s->bitrate;
1503 3 av_channel_layout_uninit(&avctx->ch_layout);
1504 3 avctx->ch_layout.order = AV_CHANNEL_ORDER_UNSPEC;
1505 3 avctx->ch_layout.nb_channels = s->nb_channels;
1506 3 avctx->sample_rate = s->rate;
1507 3 avctx->sample_fmt = AV_SAMPLE_FMT_FLT;
1508
1509
2/2
✓ Branch 0 taken 9 times.
✓ Branch 1 taken 3 times.
12 for (int m = 0; m <= s->mode; m++) {
1510 9 ret = decoder_init(s, &s->st[m], &speex_modes[m]);
1511
1/2
✗ Branch 0 not taken.
✓ Branch 1 taken 9 times.
9 if (ret < 0)
1512 return ret;
1513 }
1514
1515 3 s->stereo.balance = 1.f;
1516 3 s->stereo.e_ratio = .5f;
1517 3 s->stereo.smooth_left = 1.f;
1518 3 s->stereo.smooth_right = 1.f;
1519
1520 3 return 0;
1521 }
1522
1523 static void speex_decode_stereo(float *data, int frame_size, StereoState *stereo)
1524 {
1525 float balance, e_left, e_right, e_ratio;
1526
1527 balance = stereo->balance;
1528 e_ratio = stereo->e_ratio;
1529
1530 /* These two are Q14, with max value just below 2. */
1531 e_right = 1.f / sqrtf(e_ratio * (1.f + balance));
1532 e_left = sqrtf(balance) * e_right;
1533
1534 for (int i = frame_size - 1; i >= 0; i--) {
1535 float tmp = data[i];
1536 stereo->smooth_left = stereo->smooth_left * 0.98f + e_left * 0.02f;
1537 stereo->smooth_right = stereo->smooth_right * 0.98f + e_right * 0.02f;
1538 data[2 * i ] = stereo->smooth_left * tmp;
1539 data[2 * i + 1] = stereo->smooth_right * tmp;
1540 }
1541 }
1542
1543 3 static int speex_decode_frame(AVCodecContext *avctx, AVFrame *frame,
1544 int *got_frame_ptr, AVPacket *avpkt)
1545 {
1546 3 SpeexContext *s = avctx->priv_data;
1547 3 int frames_per_packet = s->frames_per_packet;
1548 3 const float scale = 1.f / 32768.f;
1549 3 int buf_size = avpkt->size;
1550 float *dst;
1551 int ret;
1552
1553
1/4
✗ Branch 0 not taken.
✓ Branch 1 taken 3 times.
✗ Branch 2 not taken.
✗ Branch 3 not taken.
3 if (s->pkt_size && avpkt->size == 62)
1554 buf_size = s->pkt_size;
1555
1/2
✗ Branch 1 not taken.
✓ Branch 2 taken 3 times.
3 if ((ret = init_get_bits8(&s->gb, avpkt->data, buf_size)) < 0)
1556 return ret;
1557
1558 3 frame->nb_samples = FFALIGN(s->frame_size * frames_per_packet, 4);
1559
1/2
✗ Branch 1 not taken.
✓ Branch 2 taken 3 times.
3 if ((ret = ff_get_buffer(avctx, frame, 0)) < 0)
1560 return ret;
1561
1562 3 dst = (float *)frame->extended_data[0];
1563
1/2
✓ Branch 0 taken 3 times.
✗ Branch 1 not taken.
3 for (int i = 0; i < frames_per_packet; i++) {
1564 3 ret = speex_modes[s->mode].decode(avctx, &s->st[s->mode], &s->gb, dst + i * s->frame_size, frames_per_packet - i);
1565
1/2
✓ Branch 0 taken 3 times.
✗ Branch 1 not taken.
3 if (ret < 0)
1566 3 return ret;
1567 if (avctx->ch_layout.nb_channels == 2)
1568 speex_decode_stereo(dst + i * s->frame_size, s->frame_size, &s->stereo);
1569 if (get_bits_left(&s->gb) < 5 ||
1570 show_bits(&s->gb, 5) == 15) {
1571 frames_per_packet = i + 1;
1572 break;
1573 }
1574 }
1575
1576 dst = (float *)frame->extended_data[0];
1577 s->fdsp->vector_fmul_scalar(dst, dst, scale, frame->nb_samples * frame->ch_layout.nb_channels);
1578 frame->nb_samples = s->frame_size * frames_per_packet;
1579
1580 *got_frame_ptr = 1;
1581
1582 return (get_bits_count(&s->gb) + 7) >> 3;
1583 }
1584
1585 3 static av_cold int speex_decode_close(AVCodecContext *avctx)
1586 {
1587 3 SpeexContext *s = avctx->priv_data;
1588 3 av_freep(&s->fdsp);
1589 3 return 0;
1590 }
1591
1592 const FFCodec ff_speex_decoder = {
1593 .p.name = "speex",
1594 CODEC_LONG_NAME("Speex"),
1595 .p.type = AVMEDIA_TYPE_AUDIO,
1596 .p.id = AV_CODEC_ID_SPEEX,
1597 .init = speex_decode_init,
1598 FF_CODEC_DECODE_CB(speex_decode_frame),
1599 .close = speex_decode_close,
1600 .p.capabilities = AV_CODEC_CAP_DR1 | AV_CODEC_CAP_CHANNEL_CONF,
1601 .priv_data_size = sizeof(SpeexContext),
1602 .caps_internal = FF_CODEC_CAP_INIT_CLEANUP,
1603 };
1604